Data and AI-Enabled Biological Design

Artificial intelligence models trained on large volumes of biological data (AI-bio models) have demonstrated the growing abilities to support of basic scientific research goals. But some AI-bio models may be dual use, providing both beneficial capabilities and potentially dangerous ones. A nefarious actor with access to a frontier AI-bio model might be able to use it to design a pathogen with harmful phenotypic characteristics that enhance transmissibility. But model capabilities are closely linked to the data used to train them, and much less attention has been devoted to the relationship between dangerous capabilities and biological training data. The data that are included (or excluded) in model training heavily influences the models' capabilities and limitations. Governance of data used to train AI-bio models could be a useful way to allow beneficial scientific research while safeguarding against potentially dangerous capabilities.

The authors of this paper assess current knowledge about the link between biological data and AI-bio model capabilities, describe the anticipated impacts of new biological data sources, and outline potentially dangerous capabilities that could come from broad availability of certain types of biological data. They then recommend strategies to limit the potentially dangerous capabilities arising from biological data, including options for governance of experiments and data creation, governance of curation and aggregations of data, controls on access to collections of data, and governance of the use of data for model training.